What the Dreamforce 2026 customer stories say about production agents

Salesforce says it ran 21 live customer stories at Dreamforce 2026, up from 17 last year, and most came with a number. Ninety-six percent. Ninety-five percent. Seventy-nine percent. Twenty-seven thousand employees. Sixty percent of a pipeline.

None of those figures are mine and none are audited. They are what the customers and Salesforce chose to publish. That does not make them useless, and the full recap has the announcements around them. Read as a set, they describe the shape of an agent that survives production better than any feature list did, and the gaps between them tell you what to ask before you believe the next one.

The numbers, with their sources

The TSA’s Ace agent, on Government Cloud, handles about 100,000 conversations a month and resolves 96% of routine questions without escalation, on topics like packing rules and medical devices, with a projected saving of more than 11,660 staff hours a year.

Live Nation’s Venue Agent answers parking, accessible seating, door times and event logistics across its US venues and resolves 95% of inquiries without a human. Salesforce’s Erin Oles noted the first deployment, at BottleRock, took under 30 days.

From the named-agent announcement, which I covered in what a named agent is: Engine has Casey resolving 50% of chat inquiries. Anthropic has Fin resolving 79% of conversations. Hibbett has Carter handling 90% of core shopper journeys after a six-week deployment. Asana reports four times the conversation volume with Piper and a 45-day average deployment. Perk says Hunter, still in pilot, builds 60% of its pipeline. Autism Queensland has Paige resolving 70% of administrative requests.

From the press releases: Adecco is rolling Coworker out to 27,000 employees in more than 40 countries, on top of agents live in ten countries that Adecco says cover half its revenue.

Siemens runs two agents over 2,500 inbound leads a month across 18,000 sellers and says every one becomes a productive sales conversation. Fulton Bank, per Salesforce Ben’s keynote roundup, cites 80,000 hours saved, $389 million in loans and deposits attributed to the technology, and complaint resolution four days faster.

Pattern one: the high percentages are on routine questions

The three biggest numbers, 96, 95 and 79, are resolution rates on intents that were FAQ-shaped before any agent touched them. Packing rules. Door times. Product questions. The denominator in each is a queue that a knowledge article could already answer most of, and the agent’s job is to answer it faster and around the clock.

The result is real, and it is the one to expect from a narrow agent on routine volume. The number says nothing about refund disputes, complaints or anything with a policy judgment in it. The TSA’s own wording is “routine questions”. Live Nation’s list is logistics. When you see a resolution rate, the first question is what was in the queue.

Pattern two: fast deployments are narrow deployments

Thirty days at BottleRock. Forty-five days at Asana. Six weeks at Hibbett. Each of those is one intent family on one channel with one clear rule: fan logistics, inbound qualification, guided shopping. None of them is a service desk that handles everything.

The speed comes from the scope, not from the platform being faster than platforms used to be. A team that scopes its first agent to one queue with a clear definition of done gets a number in six weeks. A team that scopes it to “customer service” gets a six-month project with no number, which is the failure pattern I described in why agent projects fail.

Pattern three: the regulated flows use a router and subagents

The story that told me most was not a percentage. Southwest Airlines showed a structured Agent Router with dedicated subagents for intake, document validation, reimbursements and ticket refunds on a customer name change. Fulton Bank’s complaint-resolution improvement sits in the same kind of flow: regulated, multi-step, with a document to check and money to move.

In both, the agent does not adjudicate. It gathers, validates and routes, and each step is a separate subagent with a narrow scope. Both follow the subagent and topic design I keep pushing.

The same shape is what Agent Optimizer can inspect, because a router gap is something a tool can find when the routes are explicit. The demo of Optimizer finding one, on refunds over $200, was the other half of the same lesson.

Pattern four: the employee rollouts are about context, not autonomy

Adecco’s 27,000 users and Fulton Bank’s sellers are not stories about an agent doing the job. They are Coworker rollouts, and Adecco’s CEO, Denis Machuel, described the value as a single interface that knows candidate history and client context at the moment a human decision matters.

Machuel’s description is the right one. Coworker answers from your data and hands the next step to an agent; the human still decides. The number that matters in those rollouts is adoption, and the risk is the permission model, because a search box that inherits every access decision you ever made will surface the ones you forgot.

Autism Queensland’s figure belongs in the same group. Paige, the IT and HR agent, resolves 70% of its administrative requests, which is an employee-facing FAQ queue with the same shape as the customer-facing ones above: routine, well documented, high volume. The pattern holds on both sides of the org chart.

What the numbers leave out

Every figure above omits at least one of the following, and the omission is usually the interesting part.

Whether resolution means the customer’s problem was solved or the conversation ended without a transfer. Containment and resolution are different numbers, and a vendor sentence rarely says which.

The cost per conversation. Not one of the 21 stories put a credit figure next to a resolution rate, and the Flex Credit arithmetic is what decides whether 95% is a good trade.

The time period and the baseline. “Four times the conversation volume” needs a before, a window, and a definition of conversation.

What “100% into productive sales conversations” measures. Siemens’ figure is about one stage of one funnel, a process claim rather than a revenue one, and the release does not say what happens next.

Escalation quality. A 96% resolution rate says nothing about how the other 4% arrived at a human: with context, or as a cold transfer. The handoff design is where that gets decided.

A checklist for reading a vendor case study

Six questions, in the order I ask them.

What was in the queue before the agent, and what share of it was routine?
Does the headline rate mean resolved, contained or deflected, and by whose definition?
How narrow was the scope, and how long did the first version take?
Where does a human enter the flow, and what arrives with the handoff?
What did it cost per conversation, and who measured it?
Would the story survive if the number were half as good? A design that still makes sense at 48% is a good design. One that only works at 96% is a marketing number.

Understanding the basics

What results did Salesforce customers report at Dreamforce 2026?

Among the published figures: the TSA resolves 96% of routine questions with its Ace agent, Live Nation resolves 95% of fan inquiries, Anthropic resolves 79% of conversations with Fin, Hibbett handles 90% of core shopper journeys with Carter, and Adecco is rolling Agentforce Coworker out to 27,000 employees. All are the customers’ or Salesforce’s own numbers.

Are the Dreamforce customer numbers independently verified?

No. They come from Salesforce press releases, Salesforce’s announcement stories and keynote coverage, and each is the customer’s or Salesforce’s claim. Read them for the shape of what works, and ask what the denominator was before comparing them to your own queue.

How long does an Agentforce deployment take?

The Dreamforce stories cite 30 days at Live Nation’s first venue, 45 days at Asana and six weeks at Hibbett. Each of those was a narrow agent on one channel with one intent family. A broadly scoped service agent takes far longer, and usually does not produce a comparable number.

What is a realistic resolution rate for an AI agent?

It depends entirely on the queue. Rates above 90% in the Dreamforce stories were all on routine, FAQ-shaped intents such as logistics and packing rules. On queues with policy judgments, refunds or complaints, expect the agent to gather and route rather than resolve, and measure escalation quality instead of a single percentage.

Read the design, not the percentage

Twenty-one stories, four patterns. Narrow scope on routine volume, fast because it was narrow, routers and subagents where the rules bite, and context rather than autonomy for employees. Every high number on stage came from an agent built that way.

Copy the design. Treat the percentage as the customer’s, and go measure your own.

If you want your first agent scoped the way the Dreamforce numbers were earned, one queue, one rule, one number, talk to us, or start with the Agentforce readiness assessment.